Heterogeneous appetite patterns in depression: computational modeling of nutritional interoception, reward processing, and decision-making

Abstract

Accurate interoceptive processing in decision-making is essential to maintain homeostasis and overall health. Disruptions in this process have been associated with various psychiatric conditions, including depression. Recent studies have focused on nutrient homeostatic dysregulation in depression for effective subtype classification and treatment. Neurophysiological studies have associated changes in appetite in depression with altered activation of the mesolimbic dopamine system and interoceptive regions, such as the insular cortex, suggesting that disruptions in reward processing and interoception drive changes in nutrient homeostasis and appetite. This study aimed to explore the potential of computational psychiatry in addressing these issues. Using a homeostatic reinforcement learning model formalizing the link between internal states and behavioral control, we investigated the mechanisms by which altered interoception affects homeostatic behavior and reward system activity via simulation experiments. Simulations of altered interoception demonstrated behaviors similar to those of depression subtypes, such as appetite dysregulation. Specifically, reduced interoception led to decreased reward system activity and increased punishment, mirroring the neuroimaging study findings of decreased appetite in depression. Conversely, increased interoception was associated with heightened reward activity and impaired goal-directed behavior, reflecting an increased appetite. Furthermore, effects of interoception manipulation were compared with traditional reinforcement learning parameters (e.g., inverse temperature β and delay discount γ), which represent cognitive-behavioral features of depression. The results suggest that disruptions in these parameters contribute to depressive symptoms by affecting the underlying homeostatic regulation. Overall, this study findings emphasize the importance of integrating interoception and homeostasis into decision-making frameworks to enhance subtype classification and facilitate the development of effective therapeutic strategies.

Keywords: appetite; computational neuroscience; computational psychiatry; decision-making; dopamine; homeostasis; homeostatic reinforcement learning.

Figure 1

Nutritional homeostatic maintenance according to the homeostatic reinforcement learning (HRL) model. (A) Schematic of the computational process of the HRL model. Herein, η represents the activity of interoceptive processing, β denotes the inverse temperature, and γ signifies delay discounting; η reflects the degree to which the difference between the internal state and the ideal state is overestimated or underestimated, characterizing the nature of interoceptive information processing until it is conveyed to the reward system. The inverse temperature β indicates the extent to which decision-making reflects the learning history. The delay discounting parameter γ is used in updating the state-action values. (B) Definition of the state and two actions in intake-after-food-restriction simulations. (C) Example of homeostatic behavior. Changes in the internal nutritional state (H), value of each action (Q-value), selected actions (a), probability of intake (P(Intake)), and magnitude of reward (R) are plotted. Solid lines indicate the results of a single trial. Dotted lines in the panel of the internal state indicate the ideal point (H^* = 200) of the nutrient. In the panel related to actions (a), action 1 indicates “intake,” and action 0 indicates “do nothing.” At the beginning of the simulation, internal nutritional state was 100, and Q-values for each action were set to 0. After several random selections of action, Q-value of nutrient intake was increased, and the internal nutritional state quickly reached the ideal point, maintaining homeostatic regulation of behavior.

Figure 2

Altered interoception in the mountain-climbing task. (A) Definitions of eight states and two actions at each state in the mountain-climbing task. (B) Time series relationships among variables. In all 15 episodes, the agent started its actions from S0. Upon reaching S7 and performing action a71 (major feeding), the episode concluded, and the agent moved to S0 to begin the next episode. The external state was reset to S0, and the internal state was initialized to 100, while the state-action values and the predicted changes in the internal state due to actions were carried over. Completing 15 episodes in this manner constituted one agent’s mountain-climbing task, which was conducted across 30 agents (30 trials). Results from performing this mountain-climbing task under different conditions, such as variations in interoceptive modulation, were compared across multiple metrics. (C) S7-rate of the control, low interoception, and high interoception models. (D) Total timesteps per trial. (E) Total number of minor intakes. (F) Trajectories of each variable in the 5th episode of the control, low interoception, and high interoception models. Significance was determined using the Student’s t-test (B,C) or Wilcoxon rank-sum test (D) (30 trials). ***p < 0.001; N.S., not significant.

Figure 3

Altered interoception in the intake-after-food-restriction task. (A) Average of reward with all intake behavior in a single episode determined from the simulations with altered interoception models. (B) Sum of punishments during one episode with altered interoception. (C) Sum of drive in one episode determined from the simulated lesion models with altered interoception. (D) Total intake in an episode with altered interoception. (E) Example of transition of variable from episodes of the control, low interoception, and high interoception models. In panels (A–D), Student’s t-test or Welch’s t-test was used for between-group comparisons after Levene’s test. **p < 0.01 and ***p < 0.001 (N = 40).

Figure 4

Alterations in reinforcement learning (RL) parameters in the intake-after-food-restriction task. (A) Average rewards per intake behavior in single episodes and sum of punishment in each episode determined from the simulations with altered inverse temperature (β). (B) Example transitions of variables of the altered β models. (C) Average rewards with all intake behavior in single episodes and sum of punishment in each episode determined from simulations with altered discount ratio (γ). (D) Example transitions of variables of the altered γ models. (E) Total number of intakes in single episodes determined from simulated lesion models with altered β. (F) Total number of intakes in single episodes determined from models with altered γ. (G) Sum of drive during single episodes of models with altered β. (H) Sum of drive during single episodes of models with altered γ. ***p < 0.001 (N = 40); N.S., not significant. In panels (A), (C), and (E–H), Student’s t-test or Welch’s t-test was used after Levene’s test.

Figure 5

Alterations in RL parameters in the mountain-climbing task. (A) Results of the control, low γ, and low β models referred to the ratios determined from the same calculations as those shown in Figure 2C. (B) The same measure of Figure 2D. (C) The same measure of Figure 2E. (D) Trajectories of each variable in the final episode of low γ and low β models. N.S., not significant; **p < 0.01 and ***p < 0.001.